FedChill: Adaptive Temperature Scaling for Federated Learning in Heterogeneous Client Environments

ICLR 2026 Conference Submission13456 Authors

Keywords: Federated Learning, Non-IID Data, Client Drift, Data Heterogeneity, Temperature Scaling

Abstract: Federated Learning (FL) enables collaborative model training with data privacy but suffers in non-i.i.d. settings due to client drift, which degrades both global and local generalizability. Recent works show that clients can benefit from lower softmax temperatures for optimal local training. However, existing methods apply a uniform value across all participants, which may lead to suboptimal convergence and reduced generalization in non-i.i.d. client settings. We propose FedChill, a heterogeneity-aware strategy that adapts temperatures to each client. FedChill initializes temperatures using a heterogeneity score, quantifying local divergence from the global distribution, without exposing private data, and applies performance-aware decay to adjust temperatures dynamically during training. This enables more effective optimization under heterogeneous data while preserving training stability. Experiments on CIFAR-10, CIFAR-100, and SVHN show that FedChill consistently outperforms baselines, achieving up to 8.35\% higher global accuracy on CIFAR-100 with 50 clients, while using 2.26$\times$ fewer parameters than state-of-the-art methods.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 13456